If this is a bsim3v3 or a bsim4 model, they are not strictly capacitances - they are partial derivatives:
cgs (F) dQg_dVs
csg (F) dQs_dVg
Many years ago (over 15) spectre changed to use the native operating point parameters in the model rather than the traditional "SPICE" meanings of cgs etc, and I wrote a note describing this change (which published by a colleague at the time as this solution). This explains a bit more why they are negative and that the traditional "cgs" capacitance value would be found by taking -csg as reported by spectre.